Hello Esri Community,
I recently purchased a high-performance workstation with current-generation hardware, expecting to run ArcGIS Pro Deep Learning tools on the GPU. However, I discovered that the Deep Learning stack in ArcGIS Pro appears to be limited by a CUDA version mismatch, making GPU acceleration unusable on my system.
My hardware / setup
ArcGIS Pro: 3.6
GPU: NVIDIA RTX PRO 2000 (Blackwell, compute capability sm_120)
OS: Windows (workstation class)
Goal: Run ArcGIS Pro Deep Learning tools (inference and training) with GPU acceleration via the required conda-based Python environment
Problem summary
ArcGIS Pro 3.6’s Deep Learning Python environment seems to ship with / pin a CUDA runtime around CUDA 12.6 (as part of the Esri deep learning frameworks/environment). In contrast, Blackwell GPUs require at least CUDA 12.8 for native support of sm_120. As a result, ArcGIS Pro Deep Learning tools cannot use the GPU (or fail with CUDA-related errors), even though the machine is otherwise fully capable.
That means CUDA/GPU acceleration is not available for ArcGIS Pro Deep Learning tools on this hardware generation. AND i didn't find any easily available information about these kinds of problems in advance !!!
What I am looking for
Is there an official or supported solution/workaround to enable GPU acceleration for Blackwell (sm_120) in ArcGIS Pro Deep Learning tools?
Is Esri planning to update the ArcGIS Pro Deep Learning frameworks/environment to CUDA 12.8+ (or otherwise add Blackwell support), and if so, is there a timeline?
If there is a recommended interim workflow (e.g., running inference/training outside Pro and importing results), what is the most practical approach from Esri’s perspective?
If needed, I can provide the exact error message text and environment/package details.
Thank you for any guidance,
rkalasek
